Journal of Machine Learning Research 16 (2015) 2589-2609 Submitted 9/14; Revised 3/15; Published 12/15 Counting and Exploring Sizes of Markov Equivalence Classes of Directed Acyclic Graphs Yangbo He heyb@pku.edu.cn Jinzhu Jia jzjia@math.pku.edu.cn LMAM, School of Mathematical Sciences, LMEQF, and Center for Statistical Science Peking University, Beijing 100871, China Bin Yu Departments of Statistics and EECS University of California at Berkeley, Berkeley, CA 94720 binyu@stat.berkeley.edu Editor: Isabelle Guyon and Alexander Statnikov Abstract When learning a directed acyclic graph (DAG) model via observational data, one generally cannot identify the underlying DAG, but can potentially obtain a Markov equivalence class. The size (the number of DAGs) of a Markov equivalence class is crucial to infer causal effects or to learn the exact causal DAG via further interventions. Given a set of Markov equivalence classes, the distribution of their sizes is a key consideration in developing learning methods. However, counting the size of an equivalence class with many vertices is usually computationally infeasible, and the existing literature reports the size distributions only for equivalence classes with ten or fewer vertices. In this paper, we develop a method to compute the size of a Markov equivalence class. We first show that there are five types of Markov equivalence classes whose sizes can be formulated as five functions of the number of vertices respectively. Then we introduce a new concept of a rooted sub-class. The graph representations of rooted subclasses of a Markov equivalence class are used to partition this class recursively until the sizes of all rooted subclasses can be computed via the five functions. The proposed size counting is efficient for Markov equivalence classes of sparse DAGs with hundreds of vertices. Finally, we explore the size and edge distributions of Markov equivalence classes and find experimentally that, in general, (1) most Markov equivalence classes are half completed and their average sizes are small, and (2) the sizes of sparse classes grow approximately exponentially with the numbers of vertices. Keywords: directed acyclic graphs, Markov equivalence class, size distribution, causality 1. Introduction Graphical models based on directed acyclic graphs (DAGs) are commonly used to derive the dependent or causal relationships in many fields such as sociology, epidemiology, and biology (Finegold and Drton, 2011; Friedman, 2004; Heckerman et al., 1999; Jansen et al., 2003; Maathuis et al., 2009). A DAG can be used to represent causal relationships of variables, where the directed edges connect the causes and their direct effects. In general, observational data is not sufficient to distinguish the underlying DAG from its statistically 2015 Yangbo He, Jinzhu Jia and Bin Yu.
He, Jia and Yu equivalent DAGs; however, it is possible to learn the Markov equivalence class that contains these equivalent DAGs (Pearl, 2000; Spirtes et al., 2001). This has led to many works that try to learn a Markov equivalence class or to learn causality based on a given Markov equivalence class from observational or experimental data (Castelo and Perlman, 2004; Chickering, 2002; He and Geng, 2008; Maathuis et al., 2009; Perlman, 2001). The size of a Markov equivalence class is the number of DAGs in the class. This size has been used in papers to design causal learning approaches or to evaluate the complexity of a Markov equivalence class in causal learning. For example, He and Geng (2008) proposes several criteria, all of which are defined on the sizes of Markov equivalence classes, to minimize the number of interventions; this minimization makes helpful but expensive interventions more efficient. Based on observational data, Maathuis et al. (2009) introduces a method to estimate the average causal effects of the covariates on the response by considering the DAGs in the equivalence class; the size of the class determines the complexity of the estimation. Chickering (2002) shows that causal structure search in the space of Markov equivalence class models could be substantially more efficient than that in the space of DAG models if most sizes of Markov equivalence classes are large. The size of a small Markov equivalence class is usually counted via traversal methods that list all DAGs in the Markov equivalence class (Gillispie and Perlman, 2002). However, if the class is large, it may be infeasible to list all DAGs. For example, as we will show later in our experiments, the size of a Markov equivalence class with 50 vertices and 250 edges could be greater than 10 24. To our knowledge, there are no efficient methods to compute the size of a large Markov equivalence class; approximate proxies, such as the number of vertices and the number of spanning trees related to the class, have been used instead of the exact size in the literature (Chickering, 2002; He and Geng, 2008; Meganck et al., 2006). Computing the size of a Markov equivalence class is the focus of this article. We first discuss Markov equivalence classes whose sizes can be calculated just through the numbers of vertices and edges. Five explicit formulas are given to obtain the sizes for five types of Markov equivalence classes respectively. Then, we introduce rooted sub-classes of a Markov equivalence class and discuss the graphical representations of these sub-classes. Finally, for a general Markov equivalence class, we introduce a counting method by recursively partitioning the Markov equivalence class into smaller rooted sub-classes until all rooted sub-classes can be counted with the five explicit formulas. Next, we also report new results about the size and edge distributions of Markov equivalence classes for sparse graphs with hundreds of vertices. By using the proposed size counting method in this paper and an MCMC sampling method recently developed by He et al. (2013a,b), we experimentally explore the size distributions of Markov equivalence classes with large numbers of vertices and different levels of edge sparsity. In the literature, the size distributions are studied in detail just for Markov equivalence classes with up to 10 vertices by traversal methods (Gillispie and Perlman, 2002). The rest of the paper is arranged as follows. In Section 2, we provide a brief review of the concept of a Markov equivalence class. In Section 3, we propose efficient algorithms to calculate the size of a Markov equivalence class. In Section 4, we study the sizes of Markov equivalence classes experimentally. We conclude in Section 5 and finally present all proofs in the Appendix. 2590
Counting and Exploring Sizes of Markov Equivalence Classes 2. Markov Equivalence Class A graph G consists of a vertex set V and an edge set E. A graph is directed (undirected) if all of its edges are directed (undirected). A sequence of edges that connect distinct vertices in V, say {v 1,, v k }, is called a path from v 1 to v k if either v i v i+1 or v i v i+1 is in E for i = 1,, k 1. A path is partially directed if at least one edge in the path is directed. A path is directed (undirected) if all edges are directed (undirected). A cycle is a path from a vertex to itself. A directed acyclic graph (DAG) D is a directed graph without any directed cycle. Let V be the vertex set of D and τ be a subset of V. The induced subgraph D τ of D over τ, is defined to be the graph whose vertex set is τ and whose edge set contains all of those edges of D with two end points in τ. A v-structure is a three-vertex induced subgraph of D like v 1 v 2 v 3. A graph is called a chain graph if it contains no partially directed cycles. The isolated undirected subgraphs of the chain graph after removing all directed edges are the chain components of the chain graph. A chord of a cycle is an edge that joins two nonadjacent vertices in the cycle. An undirected graph is chordal if every cycle with four or more vertices has a chord. A graphical model is a probabilistic model for which a DAG denotes the conditional independencies between random variables. A Markov equivalence class is a set of DAGs that encode the same set of conditional independencies. Let the skeleton of an arbitrary graph G be the undirected graph with the same vertices and edges as G, regardless of their directions. Verma and Pearl (1990) proves that two DAGs are Markov equivalent if and only if they have the same skeleton and the same v-structures. Moreover, Andersson et al. (1997) shows that a Markov equivalence class can be represented uniquely by an essential graph. Definition 1 (Essential graph) The essential graph of a DAG D, denoted as C, is a graph that has the same skeleton as D, and an edge is directed in C if and only if it has the same orientation in every equivalent DAG of D. It can be seen that the essential graph C of a DAG D has the same skeleton as D and keeps the v-structures of D. Andersson et al. (1997) also introduces some properties of an essential graph. Lemma 2 (Andersson et al. (1997)) Let C be an essential graph of D. Then C is a chain graph, and each chain component C τ of C is an undirected and connected chordal graph, where τ is the vertex set of the chain component C τ. Let SizeMEC(C) denote the size of the Markov equivalence class represented by C (size of C for short). Clearly, SizeMEC(C) = 1 if C is a DAG; otherwise C may contain more than one chain component, denoted by C τ1,, C τk. From Lemma 2, each chain component is an undirected and connected chordal graph (UCCG for short); and any UCCG is an essential graph that represents a Markov equivalence class (Andersson et al., 1997). We can calculate the size of C by counting the DAGs in Markov equivalence classes represented by its chain components using the following equation (Gillispie and Perlman, 2002; He and Geng, 2008): SizeMEC(C) = k SizeMEC(C τi ). (1) i=1 2591
He, Jia and Yu To count the size of Markov equivalence class represented by a UCCG, we can generate all equivalent DAGs in the class. However, when the number of vertices in the UCCG is large, the number of DAGs in the corresponding Markov equivalence class may be huge, and the traversal method proves to be infeasible to count the size. This paper tries to solve this counting problem for Markov equivalence classes of DAGs with hundred of vertices. 3. The Size of Markov Equivalence Class In order to obtain the size of a Markov equivalence class, it is sufficient to compute the size of Markov equivalence classes represented by undirected and connected chordal graph (UCCGs) according to Lemma 2 and Equation (1). In Section 3.1, we discuss Markov equivalence classes represented by UCCGs whose sizes are functions of the number of vertices. Then in Section 3.2.1, we provide a method to partition a Markov equivalence class into smaller subclasses. Using these methods, finally in Section 3.2.2, we propose a recursive approach to calculate the size of a general Markov equivalence class. 3.1 Size of Markov Equivalence Class Determined by the Number of Vertices Let U p,n be an undirected and connected chordal graph (UCCG) with p vertices and n edges. Clearly, the inequality p 1 n p(p 1)/2 holds for any UCCG U p,n. When U p,n is a tree, n = p 1 and when U p,n is a completed graph, n = p(p 1)/2. Given p and n, in some special cases, the size of a UCCG U p,n is completely determined by p. For example, it is well known that a Markov equivalence class represented by a completed UCCG with p vertices contains p! DAGs. Besides the Markov equivalence classes represented by completed UCCGs, there are five types of UCCGs whose sizes are also functions of p. We present them as follows. Theorem 3 Let U p,n be a UCCG with p vertices and n edges. In the following five cases, the size of the Markov equivalence class represented by U p,n is determined by p. 1. If n = p 1, we have SizeMEC(U p,n ) = p. 2. If n = p, we have SizeMEC(U p,n ) = 2p. 3. If n = p(p 1)/2 2, we have SizeMEC(U p,n ) = (p 2 p 4)(p 3)!. 4. If n = p(p 1)/2 1, we have SizeMEC(U p,n ) = 2(p 1)! (p 2)!. 5. If n = p(p 1)/2, we have SizeMEC(U p,n ) = p!. For the UCCGs other than the above five cases, it seems that the sizes of the corresponding Markov equivalence classes cannot be completely determined by the numbers of vertices and edges; the sizes of these Markov equivalence classes may depend on the exact essential graphs. Below, we display several classes of this kind for n = p + 1 or n = p(p 1)/2 3 in Example 1. Example 1. Figure 1 displays four UCCGs. Both U 5,6 and U 5,6 have 6 edges, and both U 5,7 and U 5,7 have 7 edges. We have that SizeMEC(U 5,6) = 13, SizeMEC(U 5,6 ) = 12, SizeMEC(U 5,7 ) = 14 and SizeMEC(U 5,7 ) = 30. Clearly, in these cases, the sizes of Markov equivalence classes are not completely determined by the numbers of vertices and edges. 2592
Counting and Exploring Sizes of Markov Equivalence Classes v 1 v 2 v 3 v 4 v 5 (U 5,6 ) v 1 v 2 v 3 v 4 v 5 (U 5,6 ) v 1 v 2 v 3 v 4 v 5 (U 5,7 ) v 1 v 2 v 3 v 4 v 5 (U 5,7 ) Figure 1: Examples that UCCGs with the same number of edges have different sizes. 3.2 Size of a General Markov Equivalence Class In this section, we introduce a general method to count the size of a Markov equivalence class. We have shown in Theorem 3 that there are five types of Markov equivalence classes whose sizes can be calculated with five formulas respectively. For one any other Markov equivalence class, we will show in this section that it can be partitioned recursively into smaller subclasses until the sizes of all subclasses can be calculated with the five formulas above. We first introduce the partition method and the graph representation of each subclass in Section 3.2.1. Then provide a size counting algorithm for one arbitrary Markov equivalence class in Section 3.2.2. The proofs of all results in this section can be found in the Appendix. 3.2.1 Methods to Partition a Markov Equivalence Class Let U be a UCCG, τ be the vertex set of U and let D be a DAG in the equivalence class represented by U. A vertex v is a root of D if all directed edges adjacent to v are out of v, and D is v-rooted if v is a root of D. To count DAGs in the class represented by U, below, we show that all DAGs can be divided into different groups according to the roots of the DAGs and then we calculate the numbers of the DAGs in these groups separately. Each group is called as a rooted sub-class defined as follows. Definition 4 (a rooted sub-class) Let U be a UCCG over τ and v τ. We define the v-rooted sub-class of U as the set of all v-rooted DAGs in the Markov equivalence class represented by U. The following theorem provides a partition of a Markov equivalence class represented by a UCCG and the proof can be found in Appendix. Theorem 5 (a rooted partition) Let U be a UCCG over τ = {v 1,, v p }. For any i {1,, p}, the v i -rooted sub-class is not empty and this set of p sub-classes forms a disjoint partition of the set of all DAGs represented by U. Below we describe an efficient graph representation of v-rooted sub-class. One reason to this graph representation is that for any v τ, the number of DAGs in the v-rooted sub-class might be extremely huge and it is computationally infeasible to list all v-rooted DAGs in this sub-class. Using all DAGs in which v is a root, we construct a rooted essential graph in Definition 6. Definition 6 (rooted essential graph) Let U be a UCCG. The v-rooted essential graph of U, denoted by U (v), is a graph that has the same skeleton as U, and an edge is directed in U (v) if and only if it has the same orientation in every v-rooted DAG of U. 2593
He, Jia and Yu From Definition 6, a rooted essential graph has more directed edges than the essential graph U since the root introduces some directed edges. Algorithm 3 in Appendix shows how to generate the v-rooted essential graph of a UCCG U. We display the properties of a rooted essential graph in Theorem 7 and the proof can be found in Appendix. Theorem 7 Let U be a UCCG and U (v) be a v-rooted essential graph of U defined in Definition 6. The following three properties hold for U (v) : 1. U (v) is a chain graph, 2. every chain component U (v) τ of U (v) is chordal, and 3. the configuration v 1 v 2 v 3 does not occur as an induced subgraph of U (v). Moreover, there is a one-to-one correspondence between v-rooted sub-classes and v-rooted essential graphs, so U (v) can be used to represent uniquely the v-rooted sub-class of U. From Theorem 7, we see that the number of DAGs in a v-rooted essential graph U (v) can be calculated by Equation (1) which holds for any essential graph. To use Equation (1), we have to generate all chain components of U (v). Below we introduce an algorithm called ChainCom(U, v) in Algorithm 1 to generate U (v) and all of its chain components. Algorithm 1: ChainCom(U, v) Input: U, a UCCG; v, a vertex of U. Output: v rooted essential graph of U and all of its chain components. 1 Set A = {v}, B = τ \ v, G = U and O = 2 while B is not empty do 3 Set T = {w : w in B and adjacent to A} ; 4 Orient all edges between A and T as c t in G, where c A, t T ; 5 repeat 6 for each edge y z in the vertex-induced subgraph G T do 7 if x y z in G and x and z are not adjacent in G then 8 Orient y z to y z in G 9 until no more undirected edges in the vertex-induced subgraph G T can be oriented; 10 Set A = T and B = B \ T ; 11 Append all isolated undirected graphs in G T to O; 12 return G and O We show that Algorithm 1 can generate rooted essential graph and the chain components of this essential graph correctly in the following theorem. Theorem 8 Let U be a UCCG and let v be a vertex in U. Let O and G be the outputs of Algorithm 1 given U and v. Then G is the v-rooted essential graph G = U (v) of U and O is the set of all chain components of U (v). 2594
Counting and Exploring Sizes of Markov Equivalence Classes The following example displays rooted essential graphs of a UCCG and illustrates how to implement Algorithm 1 to construct a rooted essential graph and how to generate all DAGS in the corresponding rooted sub-classes. Example 2. Figure 2 displays an undirected chordal graph U and its rooted essential graphs. There are five rooted essential graphs {U (vi) } i=1,,5. We need to construct only U (v1), U (v2) and U (v3) since U (v4) and U (v5) are symmetrical to U (v1) and U (v3) respectively. Clearly, they satisfy the conditions shown in Theorem 7. Given U in Figure 2, U (v1) is constructed according to Algorithm 1 as follows: (1) set T = {v 2, v 3 } in which vertices are adjacent to v 1, orient v 1 v 2, v 1 v 3 to v 1 v 2, v 1 v 3 respectively; (2) set T = {v 4, v 5 } in which vertices are adjacent to {v 2, v 3 }, orient v 2 v 4, v 2 v 5, v 3 v 5 to v 2 v 4, v 2 v 5, v 3 v 5 respectively; (3) orient v 5 v 4 to v 5 v 4 because v 3 v 5 v 4 occurs but v 3 and v 4 are not adjacent. By orientating the undirected edges of the chain components of a rooted essential graph with the constraint that no new v-structures and directed cycle occur, we can generate all DAGS in the corresponding sub-class (He and Geng, 2008; Meek, 1995; Verma, 1992). For example, consider U (v1) in Figure 2, we get two v 1 -rooted DAGs by orienting v 2 v 3 to v 2 v 3 or v 2 v 3. v 1 v 2 v 3 v 1 v 2 v 3 v 1 v 2 v 3 v 1 v 2 v 3 v 4 v 5 (U) v 4 v 5 (U (v1) ) v 4 v 5 (U (v2) ) v 4 v 5 (U (v3) ) Figure 2: An undirected chordal graph U and its rooted essential graphs: U (v 1), U (v 2), and U (v 3). Now we can partition a Markov equivalence class represented by a UCCG into disjoint sub-classes, each of which can be represented by a rooted essential graph. In the next section, we will show how to recursively implement these partitions until the sizes of the subclasses or their essential graphs can be calculated with the five formulas in Theorem 3. 3.2.2 Calculating the Size of a Markov Equivalence Class Let U be an undirected and connected chordal graph (UCCG) over τ. For any v τ, SizeMEC(U (v) ) denotes the number of DAGs in v-rooted sub-class of U. According to Theorem 5, the size of U can be calculated via the following corollary. Corollary 9 Let U be a UCCG over τ = {v i } i=1,,p. We have SizeMEC(U (vi) ) > 1 for i = 1,, p and p SizeMEC(U) = SizeMEC(U (vi) ). (2) i=1 This corollary shows that the size of Markov equivalence class represented by U can be calculated via the sizes of smaller sub-classes represented by {U (v i) } i=1,,p. The following example illustrates how to calculate the size of U in Figure 2. 2595
He, Jia and Yu Example 3. Consider again the undirected chordal graph U in Figure 2, SizeMEC(U) can be calculated as 5 i=1 SizeMEC(U (vi) ) according to Corollary 9. The sizes of the five subclasses represented by U (v1),, U (v5) are 2, 4, 3, 2, 3 respectively. Therefore, we can get that SizeMEC(U) = 2 + 4 + 3 + 2 + 3 = 14. According to Theorem 7, for any i {1,, p}, the v i -rooted essential graph U (v i) is a chain graph. If U (vi) is not directed, each of their isolated undirected subgraphs is a UCCG. Recall that we can calculate the size of a Markov equivalence class through its chain components using Equation (1), similarly, we can calculate the size of v i -rooted sub-class of U with its isolated UCCGs as follows. Corollary 10 Let U (vi) be a v i -rooted equivalent sub-class of U defined in Definition 6 and {U (v i) τ j } j=1,,l be the isolated undirected chordal sub-graphs of U (vi) over the vertex set τ j for j = 1,, l. We have SizeMEC(U (v i) ) = l j=1 SizeMEC(U (v i) τ j ). (3) Since {U (v i) τ j } j=1,,l are UCCGs according to Theorem 7, SizeMEC(U (v i) τ j ) can be calculated again via Equation (2) in Corollary 9 recursively. In this iterative approach, Equation (2) and Equation (3) are used alternately to calculate the sizes of equivalence classes represented by an undirected essential graph and a rooted essential graph. Now in Algorithm 2 we present an enumeration to give SizeMEC (U). Corollary 11 shows that the enumeration returns the size correctly. For any essential graph C, we can calculate the size of Markov equivalence class represented by C according to Equation (1) and Algorithm 2. Algorithm 2: SizeMEC (U) Input: U: a UCCG. Output: the size of Markov equivalence classes represented by U 1 Let p and n be the numbers of vertices and edges in U; 2 switch n do 3 case p 1 return p; 4 case p return 2p; 5 case p(p 1)/2 2 return (p 2 p 4)(p 3)!; 6 case p(p 1)/2 1 return 2(p 1)! (p 2)!; 7 case p(p 1)/2 return p!; 8 for j 1 to p do 9 {U 1,, U lj } ChainCom(U, v j ); 10 s j l j i=1 SizeMEC (U i) 11 return p i=1 s i Corollary 11 Let U be a UCCG and SizeMEC( ) be the function defined in Algorithm 2. The function SizeMEC(U) returns the size of Markov equivalence class represented by U. 2596
Counting and Exploring Sizes of Markov Equivalence Classes The complexity of calculating SizeMEC(U) via Algorithm 2 depends on the number of times this recursive function is called. Our experiments in the next section show that when the number of vertices in U is small, or when the number is large but U is sparse, our proposed approach is efficient. However, when U is large and dense, the proposed approach may be computational infeasible since calculating SizeMEC(U) via Algorithm 2 may require a very deep recursion. In the worst case, the time complexity of Algorithm 2 might be O(p!). For example, it might be extremely time-consuming to count SizeMEC(U) via Algorithm 2 when U is a UCCG with large p vertices and p(p 1)/2 3 edges. Fortunately, according to the experimental results in He et al. (2013a), the undirected and connected chordal subgraphs in sparse essential graphs with hundreds of vertices are mostly small. This implies that our approach may be valuable for size counting in most situations of causal learning based on sparse graphical models. In the next section, we demonstrate our approach experimentally and explore the size and edge distributions of Markov equivalence classes in sparse graphical models. 4. Experimental Results We conduct experiments to evaluate the proposed size counting algorithms in Section 4.1, and then to study sizes of Markov equivalence classes in Section 4.2. The main contributions of these experiments are as follows. 1. Our proposed approach can calculate the size of classes represented by a UCCG with a few vertices (p < 15) in seconds on a laptop of 2.7GHz and 8G RAM. When the number of vertices is large, our approach is also efficient for the graphs with a sparsity constraint. 2. For the essential graphs with a sparsity constraint, the sizes of the corresponding Markov equivalence classes are nearly exponential in p. This explains the result in Chickering (2002) that causal structure search in the space of Markov equivalence class models could be substantially more efficient than the search in the space of DAG models for learning sparse graphical models. 3. In the set of all Markov equivalence classes of DAGs with p vertices, most graphs are half-completed (nearly p 2 /4 edges exist) and the Markov equivalent classes represented by these graphs have small average sizes. This is the reason why all Markov equivalence classes have small average sizes (approximately 3.7 reported by Gillispie and Perlman (2002)) even though sparse Markov equivalence classes are huge. 4.1 Calculating the Size of Classes Represented by UCCGs In this section, we experimentally study the time complexity of our proposed counting algorithms for the UCCGs with a small p or with a large p but also with a sparsity constraint. All experiments are run on a laptop with Intel 2.7GHz and 8G RAM. Note that the chain components are mostly small from sparse Markov equivalence classes with hundreds of vertices (He et al., 2013a). The experimental results show that the proposed method is efficient to count the sizes of sparse Markov equivalence classes with hundreds of vertices. 2597
He, Jia and Yu Let U n p be the set of Markov equivalence classes with p vertices and n edges. The graphs in U n p are sparse if n is a small multiple of p. We generate random choral graphs in U n p as follows. First, we construct a tree by connecting two vertices (one is sampled from the connected vertices and the other from the isolated vertices) sequentially until all p vertices are connected. Then we randomly insert an edge such that the resulting graph is chordal, repeatedly until the number of edges reaches n. Repeating this procedure N times, we obtain N samples from U i p for each i [p 1, n]. We first consider the undirected chordal graphs with 5 to 13 vertices. Our experiments on U n p for any n < p(p 1)/2 3 show that it is most time-consuming to calculate the size of UCCGs when n = p(p 1)/2 3. Based on the samples from U n p where n = p(p 1)/2 3, we report in Table 1 the the maximum, the minimum and the average of the sizes of Markov equivalence classes and the time to count them. We see that the size is increasing exponentially in p and the proposed size-counting algorithm is computationally efficient for undirected chordal graphs with a few vertices. p 5 6 7 8 9 10 11 12 13 Min 14 60 312 1920 1.36e4 1.11e5 1.00e6 1.02e7 1.12e8 Size Mean 22 104 658 4508 3.27e4 2.90e5 2.96e6 2.92e7 3.57e8 Max 30 144 828 5616 4.39e4 3.89e5 3.84e6 4.19e7 4.99e8 Min 0 0 1.0e-3 5.0e-3 2.8e-2 1.7e-1 1.3 10.6 95 Time Mean 1.3e-4 4.3e-4 1.5e-3 6.8e-3 3.6e-2 2.2e-1 1.6 13.6 140 (sec.) Max 1.0e-3 1.0e-3 4.0e-3 1.3e-2 9.6e-2 6.4e-1 5.1 53.5 476 Table 1: The size of Markov equivalence class and the time to calculate it via Algorithm 2 based on 10 5 samples from U n p, where p ranges from 5 to 13 and n = p(p 1)/2 3 (the worst case for classes with p vertices). We also study the sets U n p that contain UCCGs with tens of vertices. The number of vertices p is set to 15, 20,, 100 and the edge constraint m is set to rp where r is the ratio of m to p. For each p, we consider four ratios: 2, 3, 4 and 5. The undirected chordal graphs in U rp p are sparse since r 5. Based on 10 5 samples, we report the average size and time in Table 2. We can see that when r 4, the algorithm just takes a few seconds even when the sizes are very huge; when the chordal graphs become denser (r > 4), the algorithm takes more time. Here we have to point out that the choral graphs generated in this experiment might not be uniformly distributed in the space of chordal graphs and that the averages in Table 1 and Table 2 are not accurate estimations of expectations of sizes and time. 4.2 Size and Edge Distributions of Markov Equivalence Classes In this section, we focus on the size and edge distributions of Markov equivalence classes of directed acyclic graphs. First, we generate a Markov chain on Markov equivalence classes of interest and simultaneously obtain the stationary distribution of the chain according to the methods in He et al. (2013a,b). Then, based on the stationary distribution of the chain, we reweigh the samples from the chain and further use them to calculate the distribution 2598
Counting and Exploring Sizes of Markov Equivalence Classes r p 15 20 30 40 50 60 70 80 90 100 2 7363 6.98e4 4.74e6 6.94e8 1.9e10 1.2e12 1.2e14 1.5e15 1.8e17 2.6e19 3 3.0e5 3.3e6 1.1e10 7.1e12 4.4e15 8.6e18 1.3e21 6.1e23 1.4e27 9.1e27 Size 4 2.7e6 5.4e8 6.7e12 2.8e16 3.5e19 5.9e22 5.8e25 1.3e29 1.3e38 1.5e34 5 4.9e7 6.7e9 8.3e14 5.4e18 1.1e24 2.8e26 2.3e30 4.8e33 5.6e40 3.8e40 2 3.2e-3 5.7e-3 1.2e-2 2.3e-2 0.028 0.037 0.059 0.074 0.090 0.15 3 Time 1.7e-2 3.8e-2 8.8e-2 0.15 0.17 0.27 0.42 0.53 0.75 0.86 4 (sec.) 0.19 0.43 0.72 1.37 1.51 2.16 3.35 3.64 6.14 9.03 5 2.89 7.07 7.91 17.49 50.43 82.99 90.37 95.54 127.25 213 Table 2: The average size of Markov equivalence classes and average counting time via Algorithm 2 are reported based on 10 5 samples from U pr p, where p ranges from 15 to 100. of Markov equivalence classes of interest. In Section 4.2.1, we study the size and edge distributions of Markov equivalence classes with tens of vertices, and in Section 4.2.2, we provide the size distributions of Markov equivalence classes with hundred of vertices under sparsity constraints. 4.2.1 Size and Edge Distribution of Markov Equivalence Classes In this section, we discuss the distributions of Markov equivalence classes on their sizes and number of edges. We use size distribution for the distribution on sizes of Markov equivalence classes, and edge distribution for the distribution on the number of edges. First, we consider the number of edges of Markov equivalence classes with p vertices for 10 p < 20. Then, we focus on the size and edge distribution of Markov equivalence classes with 20 vertices. Finally, we explore the size distributions of Markov equivalence classes with different numbers of edges to show how size distributions change with increasing numbers of edges. The numbers of edges in the Markov equivalence classes with p vertices range from 0 to p(p 1)/2. Based on a Markov chain with length of 10 6 for each p, we display in Table 3 the modes and 99% intervals of edge distributions of Markov equivalence classes with p vertices for 10 p < 20. The mode is the number that appears with the maximum probability, 99% interval is the shortest interval that covers more than 99% of Markov equivalence classes. The ratios that measure the fraction of 99% interval to p(p 1)/2 + 1 are also given. For example, consider the edge distribution of Markov equivalence classes with 10 vertices; we see that 99% of Markov equivalence classes have between 17 and 32 edges. The ratio is 16/46 0.348, where the number 16 is the length of the 99% interval [17, 32] and 46 is the length of the edge distribution s support [0, 45]. From the 99% intervals and the corresponding ratios, we see that the numbers of edges of Markov equivalence classes are sharply distributed around p 2 /4, and these distributions become sharper with increasing of p. This result is reasonable since the number of skeletons of essential graphs with k edges is ( p(p 1)/2) k, and the k-combination reaches maximum around k = p 2 /4. In Figure 3, we display the proportions of Markov equivalence classes with 20 vertices according to their sizes and the number of edges. Two scaled marginal distributions in the 2599
He, Jia and Yu p mode 99% interval ratio p mode 99% interval ratio 10 25 [17,32] 0.348 15 56 [44,68] 0.236 11 30 [22,39] 0.321 16 64 [51,77] 0.223 12 36 [26,45] 0.299 17 73 [59,87] 0.216 13 42 [32,53] 0.278 18 81 [66,96] 0.201 14 49 [38,60] 0.25 19 91 [75,106] 0.180 Table 3: The edge distributions of Markov equivalence classes with p vertices for 10 p < 20. The mode is the number that appears with the maximum probability, the 99% interval covers more than 99% of Markov equivalence classes, ratio is the fraction of the length of the 99% interval to the length of the support of edge distribution. planes are also shown. The black dashed line is the size distribution and the black solid line is the edge distribution of Markov equivalence classes. According to the marginal size distribution, we see that most of the Markov equivalence classes with 20 vertices have small sizes. For example, 26.89% of Markov equivalence classes are of size one; the proportion of Markov equivalence classes with size 10 is greater than 95%. We also see that the marginal edge distribution of Markov equivalences is concentrated around 100(= 20 2 /4). The proportion of Markov equivalence classes with 20 vertices and 100 edges is nearly 6%. To study how the size distribution changes with the number of edges, we consider Markov equivalence classes with 100 vertices and n edges for different n. Figure 4 displays the size distribution of Markov equivalence classes with 100 vertices and n edges for n = 10, 50, 100, 200, 400, 600, 1000, 1500, 2000 and 2500, respectively. We see that the sizes of Markov equivalence classes are very small when the number of edges is close to p 2 /4 = 2500. For example, when n (1000, 2500), the median of the sizes is no more than 4. These results shed light on why the Markov equivalence classes have a small average size (approximately 3.7 reported by Gillispie and Perlman (2002)). 4.2.2 Size Distributions of Markov Equivalence Classes with Sparsity Constraints We study Markov equivalence classes with p vertices and n vertices. The number of vertices p is set to 100, 200, 500 or 1000 and the maximum number of edges n is set to rp where r is the ratio of n to p. For each p, we consider four ratios: 1.2, 1.5, 3 and 5. The essential graphs with p vertices and rp edges are sparse since r 5. In each simulation, given p and r, a Markov chain with length of 10 6 Markov equivalence classes is generated. There are sixteen distributions, each of which is calculated with 10 6 essential graphs. We plot the four distributions for r = 1.2 in the main window, and the other 12 distributions for r = 1.5, 3, 5 in three sub-windows, respectively. In each distribution, the 95% quantiles and 99% quantiles are marked with diamonds and circles, respectively. We see that the sizes of equivalence classes are extremely large. The medians of size distributions are connected by a dashed line in Figure 5. It seems that there is a linear relationship between the logarithm of size and the number of vertices p. In other words, the size grows exponentially with p. 2600
Counting and Exploring Sizes of Markov Equivalence Classes 0.2689 0.016 0.0596 0.014 0.012 0.010 0.008 0.006 Proportions 0.004 0.002 0.000 5 10 Sizes 15 20 70 80 90 100 110 Edges 120 130 Figure 3: The surface displays the distribution of the Markov equivalence classes with 20 vertices. Two rescaled marginal distributions are shown in the planes. The black dashed line is the size distribution and the black solid line is the edge distribution of Markov equivalence classes. These results suggest that, to learn directed graphical models, a searching among Markov equivalence classes might be more efficient than that among DAGs since the number of Markov equivalence classes is much less than the number of DAGs when the graphs of interest are sparse. 5. Conclusions and Discussions In this paper, we propose a method to calculate the sizes of Markov equivalence classes. A rooted sub-class of a Markov equivalence class is introduced and the graph representation of this sub-class, called rooted essential graph, is proposed. We can partition a Markov equivalence class into smaller rooted sub-classes recursively until the sizes of all sub-classes can be obtained via five closed-form formulas. Then we explore the size and edge distributions of Markov equivalence classes. We study experimentally how size distribution changes with the number of edges and report the size distributions of Markov equivalence classes with hundreds of vertices under sparsity constraints. We find that the essential graphs with around p 2 /4 edges dominate in the set of all essential graphs with p vertices and the corresponding Markov equivalence classes have small sizes. This results in a small average size of all Markov equivalence classes with p vertices. For the sparse essential graphs with p vertices, we find that the sizes of the corresponding Markov equivalence classes are super-exponential in p. 2601
He, Jia and Yu 10 7 Size of MECs with 100 vertices 10 6 10 5 10 4 10 3 10 2 10 0 14 n=10 n=50 n=100 n=200 n=400 Size of MECs with 100 vertices 12 10 8 6 4 2 0 n=600 n=1000 n=1500 n=2000 n=2500 Figure 4: The size distributions of Markov equivalence classes with p vertices and n edges, where n = 10, 50, 100, 200, 400, 600, 1000, 1500, 2000 and 2500, respectively. 10 35 10 30 r=1.5 10 30 r=3 10 16 r=5 10 30 10 20 10 20 10 12 10 8 Size of Markov equivalence classes 10 25 10 20 10 15 10 10 10 10 0 10 10 0 10 4 0 r=1.2 10 5 0 p=100 p=200 p=500 p=1000 Figure 5: Size distributions of Markov equivalence classes with p vertices and at most rp edges. The lines in the boxes and the two circles above the boxes indicate the medians, the 95th, and the 99th percentiles respectively. To calculate the sizes of Markov equivalence classes, we provide five closed-form formulas for UCCGs with p vertices and n = p 1, p, p(p 1)/2 2, p(p 1)/2 1, and p(p 1)/2 edges respectively. As shown in Example 1, for other cases, say n = p(p 1)/2 3, the 2602
Counting and Exploring Sizes of Markov Equivalence Classes size of a Markov equivalence class is no longer determined by the number of vertices p; it depends on the structure of the corresponding UCCG and our proposed method might be inefficient when p is large. For these cases, it is of interest to develop more efficient algorithm, or formulas, to calculate the size of a general Markov equivalence class in the future work. Moreover, we use python to implement algorithms and experiments in this paper and the python package can be found at pypi.python.org/pypi/markovequclasses. Acknowledgments We are very grateful to Adam Bloniarz for his comments that significantly improved the presentation of our manuscript. We also thank the editors and the reviewers for their helpful comments and suggestions. This work was supported partially by NSFC (11101008, 11101005, 71271211), DPHEC-20110001120113, US NSF grants DMS-1107000, CDS&E- MSS 1228246, ARO grant W911NF-11-1-0114, AFOSR Grant FA 9550-14-0016, and the Center for Science of Information (CSoI, a US NSF Science and Technology Center) under grant agreement CCF-0939370. Appendix A. Proofs of Results In this section, we provide the proofs of the main results of our paper. We place the proof of Theorem 3 in the end of Appendix because this proof will use the results in Algorithm 1 and Corollary 9. Proof of Theorem 5: We first show that τ i -rooted sub-class is not empty. For any vertex τ i τ, we just need to construct a DAG D in which no v-structures occurs and all edges adjacent to v are oriented out of v. The maximum cardinality search algorithm introduced by Tarjan and Yannakakis (1984) can be used to construct D. Let p be the number of vertices in U, the algorithm labels the vertices from p to 1 in decreasing order. We first label τ i with p. As the next vertex to label, select an unlabeled vertex adjacent to the largest number of previously labeled vertices. We can obtain a directed acyclic graph D by orienting the undirected edges of U from higher number to lower number. Tarjan and Yannakakis (1984) show that no v-structures occur in D if U is chordal. Hence in D, there is no v-structure and all edges adjacent to v are oriented out of v. We have that D is a τ i -rooted equivalent DAG of U, thus τ i -rooted sub-class is not empty. To prove that the p sub-classes, τ i -rooted sub-classes for i = 1,, p, form a disjoint partition of Markov equivalence class represented by U, we just need to show that every equivalent DAG of U is in only one of p sub-classes. For any equivalent DAG of U, denoted by D, since D is a directed acyclic graph, there exists an order of its vertices such that all edges are oriented from the preceding vertices to their succeeding ones. Denoted by τ i the first vertex of this order, we have that all edges adjacent to τ i are oriented out of τ i. Clearly, D is in the τ i -rooted sub-class. 2603
He, Jia and Yu Below, we show that D is not in any other rooted sub-class. Suppose that D is also in another τ j -rooted sub-class (i j). Clearly, τ i and τ j are not adjacent. Since U is connected, we can find a shortest path L = {τ i τ k τ l τ j } from τ i to τ j with more than one edge. The DAG D is in both τ i -rooted and τ j -rooted sub-classes, so we have that v i v k and v j v l are in D. Considering all vertices in L, there must be a head to head like in D, and the two heads are not adjacent in D since L is shortest path. Consequently, a v-structure appears in D. This is a contradiction because U is an undirected chordal graph and D must be a DAG without v-structures. Proof of Theorem 7: Consider the proof of Theorem 6 in He and Geng (2008), we set the intervention variable to be v. If v is a root, Theorem 7 becomes a special case of Theorem 6 in He and Geng (2008). Proof of Corollary 9: Theorem 5 shows that for any i {1, 2,, p}, the τ i -rooted sub-class of U is not empty and these p sub-classes form a disjoint partition of Markov equivalence class represented by U. This implies Corollary 9 directly. Proof of Corollary 10: Since {U (v i) τ j } j=1,,l are l isolated undirected chordal sub-graphs of U (vi), the orientations of the undirected edges in a component is irrelevant to the other undirected components. This results in Equation (3) follow directly. Proof of Theorem 8: We first introduce the following Algorithm 3 that can construct a rooted essential graph. Algorithm 3: Find the v-rooted essential graph of U Input: U: an undirected and connected chordal graph; v: a vertex of U. Output: the v-rooted essential graph of U 1 Set H = U; 2 for each edge v in U do 3 Orient v to v in H. 4 repeat 5 for each edge y z in H do Rule 1: if there exists y z in H, and and z are not adjacent then Orient y z to y z in H Rule 2: if there exists y z in H then Orient y z to y z in H 6 until no more undirected edges in H can be oriented; 7 return H A similar version of Algorithm 3 is used in He and Geng (2008) to construct an essential graph given some directed edges. From the proof of Theorem 6 in He and Geng (2008), we have that the output of Algorithm 3, H, is the v-rooted essential graph of U, that is, H = U (v). Moreover, according to Theorem 7, we have that a v-rooted essential graph is a 2604
Counting and Exploring Sizes of Markov Equivalence Classes chain graph and its isolated undirected subgraphs are chain components. From Algorithm 1, we know that O contains all isolated undirected subgraphs of G. To prove that the output G of Algorithm 1 is the v-rooted essential graph U (v) of U and O is the set of all chain components of U (v), we just need to show that G = H given the same U and v. By comparing Algorithm 1 to Algorithm 3, we find that in Algorithm 1, Rule 1 that is shown in Algorithm 3 is used repeatedly, and in the output G of Algorithm 1, undirected edges can no longer be oriented by the Rule 1. If we further apply Rule 2 in Algorithm 3 to orient undirected edges in G until no undirected edges satisfy the condition in Rule 2. Denote the output as G. Clearly, the output G is the same as H obtained from Algorithm 3, that is, G = H. Therefore, to show G = H, we just need to show G = H, that is, the condition in Rule 2 does not hold for any undirected edge in G. In Algorithm 1, we generate a set T in each loop of while and the sequence is denoted by {T 1,, T n }. Setting T 0 = {v}, we have five facts as following Fact 1 All undirected edges in G occur in the subgraphs over T i for i = 1,, n. Fact 2 All edges in G between T i and T i+1 are oriented from T i to T i+1 for i = 0,, n 1. Fact 3 There is no edge between T i and T j if i j > 1. Fact 4 There are no v-structures in G. Fact 5 there is no directed cycle (all edges are directed) in G. Suppose there exist three vertices x, y and z such that both y x z and y z occur in G. Then a contradiction is implied. Since y z occurs in G, from Fact 1, there exists a set, denoted as T i containing both y and z. Moreover, y x z occurs in G, from Fact 2 and Fact 3, we have that x T i. Next we show that x, y and z have the same parents in T i 1. First, y and z have the same parents in T i 1 ; otherwise y z will be oriented to a directed edge. Denote by P 1 the same parents of y and z in T i 1 and by P 2 the parents of x in T i 1. Second, for any u P 1, if u is not a parent of x, then z x in U will be oriented to z x in G according to Algorithm 1. We have that u is also a parent of x and consequently, P 1 P 2. Third, For any u P 2, u must be a parent of y according to Fact 4. We have that P 2 P 1, and finally P 2 = P 1. We get that neither y x nor x z is oriented by any directed edge u y or u x with u T i 1 since P 2 = P 1. Let u 1 T i and u 1 y be the directed edge that orients y x in U to y x in G. Clearly, u 1 y occurs in G, and u 1 and x are not adjacent. Since y z is not directed in G, u 1 z must occur in U. Moreover, x z occurs in G and u 1 and x are not adjacent, we have that u 1 z will be oriented to u 1 z in G. Clearly, there occurs a directed cycle u 1 y x z u 1 in G. This is a contradiction according to Fact 5. We have that the condition of Rule 2 does not hold for any undirected edge in G, and consequently, G = H holds. Proof of Corollary 11: According to Corollary 9, Corollary 10, and Theorem 8, the output is the size of Markov equivalence class represented by U. 2605
He, Jia and Yu Proof of Theorem 3: Proof of (1): For a UCCG U p,n, if n = p 1, then the graph U p,n is a tree. For any vertex in U p,n, we have that U p,n (v) is a DAG according to Algorithm 1. Thus SizeMEC(U p,n) (v) = 1. Then, according to Corollary 9, SizeMEC(U p,n ) = p. Proof of (2): For a UCCG U p,n, if n = p, then the graph U p,n has one more edge than tree. Because U p,n is chordal, a triangle occurs in U p,n. For any vertex v in U p,n, we have that SizeMEC(U p,n) (v) = 2. Consequently, we have that SizeMEC(U) = 2p according to Corollary 9. Proof of (3): Let v 1,, v p be the p vertices of U p,n. There are only two pairs of vertices that are nonadjacent since p(p 1)/2 2 edges appear in U p,n. We first prove that these two pairs have a common vertex. Suppose v i v j and v k v l do not occur in U p,n and v i, v j, v k, v l are distinct vertices. Consider the subgraph induced by v i, v j, v k, v l of U p,n. Clearly, the cycle v i v k v j v l v i occurs in the induced graph and U p,n is not a chordal graph. We have that the missing two edges in U p,n are like v 1 v 2 v 3. According to Corollary 9, we have that SizeMEC(U p,n ) = p i=1 SizeMEC(U (v i) p,n ). We first consider U (v 1) p,n. All edges adjacent to v 2 in U (v 1) p,n are oriented to directed edges whose arrow is v 2 according to Algorithm 1 since v 2 is a neighbor of all neighbors of v 1, and v 1, v 2 are not adjacent in U (v 1) p,n. Removing v 2 from U (v 1) p,n, we have that the induced graph over v 1, v 3,, v p is a completed graph. This implies that the induced graph over v 3,, v p is an undirected completed graph with p 2 vertices. Therefore, we have SizeMEC(U (v 1) p,n ) = (p 2)!. Similarly, we can get that SizeMEC(U (v 3) p,n ) = (p 2)! since v 1 and v 3 are symmetric in U p,n. Consider U (v 2) p,n, according to Algorithm 1, for any vertex v j other than v 1 and v 3, we have that v 2 v j, v j v 1 and v j v 3 occur in U (v 2) p,n, and all other edges in U (v 2) p,n are undirected. Therefore, there are two isolated chain components in U (v 2) p,n, one contains the edge x 1 x 3 and the other is the subgraph induced by v 4,, v p. We have the size of first chain component is 2 and the second is (p 3)! since it is a completed graph with p 3 vertices. According to Corollary 10, SizeMEC(U (v 2) p,n ) = 2(p 3)!. We now consider U (v 4) p,n. According to Algorithm 1, to construct U (v 4) p,n, we first orient the undirected edges adjacent to v 4 in U p,n to directed edges out of v 4. Since v 4 is adjacent to all other vertices in U p,n, there are no subgraphs like v 4 v i v j with v 4 and v j nonadjacent. This results in the chain component of U (v 4) p,n being a graph with p 1 vertices and (p 1)(p 2)/2 2 edges (only v 1 v 2 v 3 missing). We have that SizeMEC(U (v 4) p,n ) = SizeMEC (U p 1,(p 1)(p 2)/2 2 ). 2606
Counting and Exploring Sizes of Markov Equivalence Classes Similarly, we can get that SizeMEC(U (v i) ) = SizeMEC (U p 1,(p 1)(p 2)/2 2 ) for any i 4 since exchanging the labels of these vertices will not change U. Therefore, we have proved the following formula SizeMEC (U p,n ) = (p 3)SizeMEC (U p 1,(p 1)(p 2)/2 2 ) + 2(p 2)! + 2(p 3)!, Finally, we show that SizeMEC (U p,n ) = (p 2 p 4)(p 3)! satisfies the formula and initial condition. First, we have SizeMEC (U 4,4 ) = (16 8) 1 = 8. Suppose SizeMEC (U p,n ) = (p 2 p 4)(p 3)! holds for p = j 1, SizeMEC (U j,j(j 1)/2 2 ) = (j 3)SizeMEC (U j 1,(j 1)(j 2)/2 2 ) + 2(j 2)! + 2(j 3)! = (j 3){[(j 1) 2 (j 1) 4][(j 1) 3]!} + 2(j 2)! + 2(j 3)! = [(j 1) 2 (j 1) 4 + 2(j 2) + 2](j 3)! = (j 2 j 4)!(j 3)! As a result, SizeMEC (U p,p(p 1)/2 2 ) = (p 2 p 4)(p 3)! holds for p = j. Proof of (4): From the condition, only one pair of vertices, denoted by v and u, is not adjacent in U p,n. Consider a v-rooted equivalence sub-class, all undirected edges adjacent to u are oriented to directed edges with arrows pointing to u, and all other edges can be orientated as a completed undirected graph. We have that SizeMEC(U p,n) (v) = (p 2)!. Similarly, we have that SizeMEC(U p,n) (u) = (p 2)!. For any vertex w other than v and u, consider any DAG in the w-rooted equivalent sub-class, all edges adjacent to w are oriented away from w, and all other edges form a new chain component with p 1 vertices and (p 1)(p 2)/2 1 edges. Consider SizeMEC(U p,p(p 1)/2 1 ) as a function of p, denoted by f(p). When p = 3, we have f(3) = 3. Hence we have following formula: Now, we show that f(p) = (p 2)f(p 1) + 2((p 2)!). f(p) = 2(p 1)! (p 2)! satisfies the formula and initial condition. First, we have f(3) = 2 2 1 = 3. Suppose f(p) = 2(p 1)! (p 2)! holds for p = j 1, f(j) = (j 2)f(j 1) + 2(j 2)! = (j 2)(2(j 2)! (j 3)!) + 2(j 2)! = 2(j 2)(j 2)! (j 2)! + 2(j 2)! = (j 2)!(2j 3) = 2(j 1)! (j 2)! As a result, f(p) = 2(p 1)! (p 2)! holds for p = j. Proof of (5): If U is an undirected and connected graph with p vertices, and p(p 1)/2 edges, then the graph is a complete graph. There are p! DAGs in the Markov equivalence class. 2607